AI guardrails are leaving defenders behind

Over the last two years, we've seen exponential growth in AI's cyber capabilities, both from Silicon Valley labs and from foreign ones.

Safeguards are not safety

AI companies frequently talk about their safeguards and how they're keeping everyone safe. So companies frequently turn to models made by frontier labs, whose cyber capabilities have been significantly cut back, and assume they're safe when the model produces nothing.

If a model that doesn't have advanced cyber capabilities is marketed as one, or is in any way presented as a cyber expert, people will have a false sense of security.

Gatekeepers for a technology they don't own

Frontier labs are acting as the gatekeepers for a technology that isn't unique to them, as we've seen from advanced foreign models. Even in advanced cyber capability programs, overreaching guardrails can block legitimate security work at a level that isn't sustainable.

We need to expand cyber capabilities to match what better-financed bad actors can bring to bear. If I have the resources to run a massive DeepSeek or Kimi model with advanced red team capabilities, and your organization isn't one of the lucky few deemed worthy by the labs, you don't stand a chance.

Massive zero-days have kept dropping over the last few weeks, even after assurances from OpenAI and Anthropic that all the important parties have access to their programs (Daybreak Blue and Project Glasswing). This week it was php2shell, two critical PHP flaws whose details leaked before the fix was out.

A force multiplier, not an expert

AI in its current guardrailed state isn't a replacement for a human expert, but it is a force multiplier. It has the potential to be more than what it is. To get there, we need oversight of these programs from an independent body, and we need to let more independent, vetted researchers and organizations access critical security tools like these.

All in all, just some frustration I'm expressing. All opinions are my own. - L